Shadowing · AI app
An AI English coach that talks, corrects and tracks your progress
Shadowing is an app Spanish speakers install on their phone. It brings speaking, listening, grammar and review practice together in one place, with natural voices.
- You: “I have 25 years”
- Correction: “I am 25 years old”Saved to your review
- Pronunciation: word by word
The problem it solves
Practicing spoken English usually takes an available teacher or several paid apps. Shadowing offers conversation, correction and progress tracking in a single app, and it runs entirely on the free tiers of AI providers.
What it automates
It automates the work of a tutor. It holds a voice conversation, corrects on the spot, scores pronunciation and builds the review from each person’s mistakes. Nobody schedules classes or grades exercises by hand.
What it does
Live voice conversation
A real call with a character that has its own voice, accent and tone, and corrects you as you talk.
Pronunciation feedback
It records your voice and marks which words came out right and which didn’t.
Shadowing
It draws your voice over the native speaker’s to compare intonation, rhythm and speed.
Daily missions
Four short tasks a day, with a streak, a points goal and a weekly report.
Spaced repetition
It saves your mistakes and builds a weekly lesson from them.
Vocabulary by profession
Topics and conversations for accounting, sales, technology, health or tourism.
How it’s built
- 01
Installable app
React and Vite as a PWA: it installs from the browser, stores progress on the device and works offline.
- 02
Several models at once
Each request goes to several models in parallel and the first valid answer is used. If one fails, the others cover.
- 03
Real-time voice
The browser connects straight to Gemini Live with a single-use temporary token. The key never leaves the server.
- 04
Tested end to end
198 unit tests and 32 automated tests on mobile and desktop, with a simulated microphone.
AI models
- Gemini 3 Flash, 3.1 Flash-Lite, 2.5 Flash and 2.5 Flash-Lite
- Answers, corrections and scoring. They race in parallel and the first valid answer wins.
- Gemini Live
- Real-time voice conversation.
- Gemini TTS
- Natural voices, with 30 voices and several accents.
- Llama 3.3 70B, GPT-OSS 120B and Qwen 3 32B
- Groq and Cerebras models that join the same race.
- Gemma 3 27B and Mistral Small
- Fallback when the others fail.
- Whisper Large v3 Turbo
- Transcribes speech in browsers without built-in recognition.
- Kokoro 82M
- A voice that runs on the device itself, offline.
- Azure Pronunciation Assessment
- Pronunciation score per phoneme.
Technologies
- React
- Vite
- PWA
- Groq
- Cerebras
- Azure Speech
- Upstash Redis
- Vercel
- Vitest
- Playwright
From this project to your operation
What it means for your company
The same foundation works for voice agents that answer calls, AI-based internal training and apps that keep running even when a provider fails.
Book a free call
30 minutes, no commitment. We send your proposal within 48 hours.